[Git-on-my-level]

New pool delegations on m5 were skipping the sooner-resetting codex-2 account because historical launch estimates were added again after provider observations had already measured the completed work. The shared selection engine could similarly strand Claude and Cursor allowance.

Use each usage window’s observation to retire completed-work debits, retain reservations for running and newly completed work, and retain long-running jobs by completion time. Calculate limiting pressure per window, choose the earlier reset on equal pressure, and include the exact documented 5% headroom boundary at both default and custom thresholds. Preserve idempotent replay, explicit model selection, account launcher ownership, and Multica authority semantics.

Ship schemas, help, the portable skill, embedded distribution and all nested bundle hashes together. Dated investigation and sanitized m5 evidence: docs/evidence/2026-11-15-capacity-routing.md.

Validation: focused routing tests, focused command/capacity race tests, independent replay/authority/headroom probes, and distribution/install/public-default/supervisor/schema checks passed on m5. The candidate selected codex-2 for implement and review, and all six configured pools planned successfully without capacity warnings. Broad m5 runs exposed environment-dependent fixture failures; the PR’s GitHub checks are the final full CI gate. Release publication and installed-runtime verification remain before acceptance.

[cursor]

PR Summary

Medium Risk
Changes how every configured capacity pool picks accounts at delegate launch; incorrect logic could misroute quota-sensitive work, though behavior is heavily regression-tested and does not touch auth or journal writes on plans.

Overview
Capacity pool selection no longer double-counts finished delegations after a provider has already reflected them in published usage facts.

The journal scan now feeds Launches with completion times (not creation-time per-account counts) into capacity.Select. Launch pressure is debited only for running work or completions newer than each window's observed_at (falling back to generated_at), with per-window limiting pressure, reset turnover clearing old-window debits, and a stable inclusive 5% headroom check via math.Nextafter.

Help text, schemas, the portable skill, embedded distribution hashes, and delegation docs are aligned with the new policy. Regression tests cover observation republishing, long-running jobs, and end-to-end pool planning; release notes and m5 routing evidence are added under docs/evidence/.

^{Reviewed by Cursor Bugbot for commit 9cba0a9. Configure here.}

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: a30ff0e1-49fa-4061-ad69-d4eb08aa0736)

[devin-ai-integration]

✅ Devin Review: No Issues Found

Devin Review analyzed this PR and found no bugs or issues to report.

[cursor]

Bugbot couldn't run - usage limit reached

Bugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit.

A user or team admin can review and increase usage limits in the Cursor dashboard.

(requestId: ca1b264e-8c73-42da-865b-d23238b1cfb1)